Methods for Automatic Groove Identificatoin in 3D Bullet Land Scans

Background: Statistics Applied to Forensic Science

• Validity of forensic evidence analysis has come under fire in the last decade
  
• 2016 PCAST report on feature-comparison methods
  
• NIST Center of Excellence established in 2016

Background: Bullet Lands

• When a gun is fired, the bullet is propelled forward through the barrel
  
• As it travels down the barrel, it makes contact with parts of the barrel
  
• Striations result from this contact
  
• Striations are observed on land engraved areas (lands). Lands are separated by groove engraved areas (grooves).

Background: Collection of Land Data

• Sensofar Confocal Light Microscope
  
• Hamby set 44 - 35 bullets from 10 consecutively rifled Ruger barrels
  • Rescanned with CSAFE’s microscope
    
  • To use as ‘base’ set; methods to be tested on several other data sets
  • Have manual groove identifications for this whole set
    
  • Each pixel: .645 square microns
  • Each land is 2mm (2000 microns) wide
• A scan of one bullet (6 individual lands) takes ~1 hour

Background: Land Surface

Background: Importance of Groove Removal

• Bullet matching algorithm
  
• Removing the underlying curved structure of land
• Looking at remaining residuals
  • Deviations from the natural curve of the land

Background: Importance of Groove Removal

Robust LOESS

Groove Identification Process

• Once we have the residuals, how do we decide where the cutoffs should be?
  
• Ad hoc approach: define a cutoff
  • 2*MAR (median absolute residual)

Comparing approaches

• Difficult to determine a metric for accuracy
  • Number incorrectly identified doesn’t work
    
• We will look at all the residuals in the areas between our predicted grooves and the manually identified grooves
  
• For each land in the data set (208 total), sum up these residuals

Comparing approaches

• The sums shown previously are calculated for each land and each method
• Then, we compare the distributions of those values

Next steps

• Moving away from ad hoc cut-off value
  
• Two-class classification procedures for individual points within the data
  • Using residuals as predictor in model
    
  • Unbalanced response

Acknowledgments

• All work was sponsored by CSAFE (Center for Statistics and Applications in Forensic Evidence), a NIST Center of Excellence
  
• Work advised by Drs. Heike Hofmann and Ulrike Genschel of CSAFE/Iowa State University